LSR2 analysis: Mechanisms through which exercise reduces symptom severity and/or functional impairment in posttraumatic stress disorder (PTSD)
Francesca Tinsdeall, Fiona Ramage, Virginia Chiocchia and Malcolm Macleod
2024-02-20
- 1. Flow of study selection and descriptives
- 2 Exercise v Control
- 3 Effects of model induction (SPS)
- 4. Observed relationships between different outcomes measures in the same cohorts of animals
- 5. Attrition bias and adverse effects of treatment
- 6. Summary of the evidence
- 7. Software used
Date report generated: 2024-04-29
1. Flow of study selection and descriptives
The flow of study selection is shown in Figure 1. Studies included were published between 2014 and 2024. Overall, this analysis includes 14 studies containing 360 comparisons.
Figure 1
Table 1 below gives a summary of the included studies for the effect of exercise interventions. N represents an aggregate of animals contributing to outcomes reported from control and treatment groups, and if the same control group has contributed to more than one experiment, it will be counted twice.
| Study | Strain | Intervention | Outcome | N |
|---|---|---|---|---|
| AMIRI, 2021 | Wistar (rat) | exercise, 28 Days | Other behavioural | 112 |
| ~ | ~ | ~ | Other neurobiological | 28 |
| AMIRI, 2022 | Wistar (rat) | exercise, 28 Days | Fear memory | 20 |
| KOYUNCUOGLU, 2021 | Wistar (rat) | exercise, 42 Days | Freezing | 24 |
| ~ | ~ | ~ | Other behavioural | 120 |
| ~ | ~ | ~ | Other neurobiological | 240 |
| MIRJALILI, 2022 | Wistar (rat) | exercise, 28 Days | BDNF | 80 |
| ~ | ~ | ~ | Other behavioural | 80 |
| ~ | ~ | ~ | Other neurobiological | 160 |
| ~ | ~ | ~ | Stress response | 40 |
| MOHAMMADI, 2024 | Wistar (rat) | exercise, 28 Days | BDNF | 20 |
| ~ | ~ | ~ | Fear memory | 20 |
| ~ | ~ | ~ | Freezing | 20 |
| ~ | ~ | ~ | Locomotor | 60 |
| ~ | ~ | ~ | Other neurobiological | 100 |
| PATKI, 2014 | Wistar (rat) | exercise, 14 Days | Locomotor | 80 |
| ~ | ~ | ~ | Other behavioural | 80 |
| ~ | ~ | ~ | Stress response | 20 |
| SHAFIA, 2017 | Wistar (rat) | exercise, 28 Days | BDNF | 20 |
| ~ | ~ | ~ | Fear memory | 56 |
| ~ | ~ | ~ | Other behavioural | 98 |
| ~ | ~ | ~ | Other neurobiological | 40 |
| ~ | ~ | ~ | Stress response | 14 |
| SHAFIA, 2019 | Wistar (rat) | exercise, 3 Days | BDNF | 80 |
| ~ | ~ | ~ | Other behavioural | 40 |
| SHAFIA, 2022 | Wistar (rat) | exercise, 28 Days | Locomotor | 80 |
| ~ | ~ | ~ | Other neurobiological | 20 |
| ~ | ~ | ~ | Stress response | 20 |
| SHAFIA, 2023a | Wistar (rat) | exercise, 28 Days | BDNF | 28 |
| ~ | ~ | ~ | Fear memory | 14 |
| ~ | ~ | ~ | Other behavioural | 126 |
| SHAFIA, 2023b | Wistar (rat) | exercise, 28 Days | BDNF | 112 |
| ~ | ~ | ~ | Other behavioural | 112 |
| ~ | ~ | ~ | Stress response | 56 |
| SHAFIA, 2023c | Wistar (rat) | exercise, 28 Days | Other behavioural | 28 |
| ~ | ~ | ~ | Other neurobiological | 70 |
| YAKHKESHI, 2022 | Wistar (rat) | exercise, 28 Days | BDNF | 40 |
| ~ | ~ | ~ | Other behavioural | 180 |
| ~ | ~ | ~ | Other neurobiological | 80 |
| ~ | ~ | ~ | Stress response | 20 |
| ZHANG, 2020 | Sprague-dawley (rat) | exercise, 28 Days | BDNF | 20 |
| ~ | ~ | ~ | Freezing | 20 |
| ~ | ~ | ~ | Locomotor | 40 |
| ~ | ~ | ~ | Neurotransmitter levels | 200 |
| ~ | ~ | ~ | Other behavioural | 20 |
| ~ | ~ | ~ | Other neurobiological | 60 |
Table 2 below gives a summary of the included studies for the effect of model induction. N represents an aggregate of animals contributing to outcomes reported from control and treatment groups, and if the same control group has contributed to more than one experiment, those animals will be counted more than once.
| Study | Strain | Outcome | N |
|---|---|---|---|
| AMIRI, 2021 | Wistar (rat) | Other behavioural | 112 |
| AMIRI, 2022 | Wistar (rat) | Fear memory | 20 |
| KOYUNCUOGLU, 2021 | Wistar (rat) | Freezing | 12 |
| ~ | ~ | Other behavioural | 60 |
| ~ | ~ | Other neurobiological | 120 |
| MIRJALILI, 2022 | Wistar (rat) | BDNF | 80 |
| ~ | ~ | Other behavioural | 80 |
| ~ | ~ | Other neurobiological | 160 |
| ~ | ~ | Stress response | 40 |
| MOHAMMADI, 2024 | Wistar (rat) | BDNF | 20 |
| ~ | ~ | Fear memory | 20 |
| ~ | ~ | Freezing | 20 |
| ~ | ~ | Locomotor | 60 |
| ~ | ~ | Other neurobiological | 100 |
| PATKI, 2014 | Wistar (rat) | Locomotor | 80 |
| ~ | ~ | Other behavioural | 80 |
| ~ | ~ | Stress response | 20 |
| SHAFIA, 2017 | Wistar (rat) | BDNF | 20 |
| ~ | ~ | Fear memory | 56 |
| ~ | ~ | Other behavioural | 98 |
| ~ | ~ | Other neurobiological | 40 |
| ~ | ~ | Stress response | 28 |
| SHAFIA, 2019 | Wistar (rat) | BDNF | 80 |
| ~ | ~ | Other behavioural | 40 |
| SHAFIA, 2022 | Wistar (rat) | Locomotor | 80 |
| ~ | ~ | Other neurobiological | 20 |
| ~ | ~ | Stress response | 20 |
| SHAFIA, 2023a | Wistar (rat) | BDNF | 28 |
| ~ | ~ | Fear memory | 14 |
| ~ | ~ | Other behavioural | 70 |
| SHAFIA, 2023b | Wistar (rat) | BDNF | 56 |
| ~ | ~ | Other behavioural | 56 |
| ~ | ~ | Stress response | 28 |
| SHAFIA, 2023c | Wistar (rat) | Other behavioural | 28 |
| ~ | ~ | Other neurobiological | 70 |
| YAKHKESHI, 2022 | Wistar (rat) | BDNF | 40 |
| ~ | ~ | Other behavioural | 100 |
| ~ | ~ | Other neurobiological | 80 |
| ~ | ~ | Stress response | 20 |
| ZHANG, 2020 | Sprague-dawley (rat) | BDNF | 20 |
| ~ | ~ | Freezing | 20 |
| ~ | ~ | Locomotor | 40 |
| ~ | ~ | Neurotransmitter levels | 200 |
| ~ | ~ | Other behavioural | 20 |
| ~ | ~ | Other neurobiological | 60 |
References of included studies are located in the appendix. Included studies used 14 unique disease model induction procedures.
1.1 Description of experiment types and methodological approach
Within the literature we identified distinct categories of experiments and the data presented would allow several meta-analytical contrasts to be drawn:
Treatment vs control. These were experiments investigating the effect of performing exercise, reported in 170 experiments from 14 publications.
In these studies the:
Control group is a group of animals that is (1) subjected to a PTSD model induction paradigm and (2) administered a control treatment (vehicle) or no treatment.
Intervention group is a group of animals that is (1) subjected to a PTSD model induction paradigm and (2) performing exercise.
Sham group is a group of animals that is (1) not subjected to a PTSD model induction paradigm and (2) administered a control treatment (vehicle) or no treatment. These data are required to allow a ‘normalised mean difference’ effect size to be calculated, given by
\[ \frac{{\bar{\mu}_C - \bar{\mu}_T}}{{\bar{\mu}_C - \bar{\mu}_S}} \times 100 \]
where \(\bar{\mu}_C\), \(\bar{\mu}_T\), \(\bar{\mu}_S\) are the mean reported scores in the control, treatment, and sham groups respectively.
Effects of disease modelling. These are experiments investigating the effect of models of PTSD, reported in 143 experiments from 14 publications.
In these studies the:
Control group is a group of animals that is (1) not subjected to a PTSD model induction paradigm and (2) is administered a control treatment (vehicle) or no treatment.
Intervention group is a group of animals that is (1) subjected to a PTSD model induction paradigm and (2) is administered a control treatment (vehicle) or no treatment.
Outcomes with ≥2 independent effect sizes were considered for meta-analysis. In this iteration of the review, this includes other behavioural, other neurobiological, bdnf, stress response, fear memory, locomotor and freezing.
All analyses were conducted allowing for the following hierarchical levels in a random effects model, which accounts for features common to experimental contrasts such as a shared control group:
Level 1: Rodent strain - effect sizes measured across experiments using the same rodent strain.
Level 2: Study - effect sizes measured from different experiments presented in the same publication.
Level 3: Experiment - effect sizes measured in the same experiment within a study, where often a control group contributes to several effect sizes.
Each level for the hierarchy was only included in the model if more than 4 categories were present for at least one of these levels. Where more than 4 categories are not present for all levels, the variance attributable to levels with fewer than 5 categories is reported as zero.
The hierarchical grouping may therefore be considered thus: Strains of laboratory animals are included in several Studies, each of which can report one or more Experiments, and each Experiment is comprised of at least two Cohorts which are considered identical except for differing in the experimental manipulation (the Intervention) or not being exposed to the disease modelling procedures (a Sham cohort, these only being used to provide a baseline for outcome measures to allow Normalised Mean Difference meta-analysis). An Experiment can include several experimental contrasts, for instance where different doses of drugs are compared to the same control group. Several outcomes can be measured and reported from the same cohort of animals.
We constructed multilevel models without Hartung-Knapp adjustments as
these are not available for rma.mv class objects in the metafor package.
Instead, the model is set to test = "t" to use t- and
F-distributions for making inferences, and dfs="contain" to
improve the method of approximating degrees of freedom of these
distributions.
The scales and units used to measure outcomes in preclinical studies often differ between studies although they may measure the same underlying biological construct. The primary effect size used for meta-analysis of preclinical studies is therefore the standardised mean difference (SMD, Hedge’s g). For experiments testing the effects of interventions we also present a sensitivity analysis using normalised mean difference (NMD), where there are sufficient data for sham procedures to allow this. This analysis is not possible for studies of the effect of single prolonged stress (SPS).
2 Exercise v Control
14 studies (170 comparisons) investigated the effects of exercise versus control. The number of studies and individual effect sizes for each outcome were:
Locomotor outcomes: 4 studies and 13 comparisons in 2 strains
Fear Memory 4 studies and 7 comparisons in 1 strain
Freezing: 3 studies and 4 comparisons in 2 strains
Other behaviours: 11 studies and 50 comparisons in 2 strains
BDNF: 8 studies and 24 comparisons in 2 strains
Neurotransmitter levels: 1 studies and 10 comparisons in 1 strain
Biochemical stress responses: 6 studies and 10 comparisons in 1 strain
Other neurobiological outcomes: 9 studies and 52 comparisons in 2 strains
2.1 Outcome 1: Locomotor activity
Multilevel analysis is only performed if there are 5 levels or more for at least one of Strain, Study and Experiment, and that is not the case here. 13 experimental comparisons were reported in 4 experiments from 4 publications and involving 2 different animal strain(s). We provide a conventional random effects model to illustrate the data. No subgroup analysis is performed. We also now include these locomotor outcomes under the ‘other behavioural’ heading below.
2.2 Outcome 2: Fear memory
Multilevel analysis is only performed if there are 5 levels or more for at least one of Strain, Study and Experiment, and that is not the case here. 7 experimental comparisons were reported in 4 experiments from 4 publications and involving 1 different animal strain(s). We provide a conventional random effects model to illustrate the data. No subgroup analysis is performed. We also now include these fear memory outcomes under the ‘other behavioural’ heading below.
2.3 Outcome 3: Freezing
Multilevel analysis is only performed if there are 5 levels or more for at least one of Strain, Study and Experiment, and that is not the case here. 4 experimental comparisons were reported in 3 experiments from 3 publications and involving 2 different animal strain(s). We provide a conventional random effects model to illustrate the data. No subgroup analysis is performed. We also now include these Freezing outcomes under the ‘other behavioural’ heading below.
2.4 Outcome 4: Other behavioural outcomes
2.4.1 Risks of bias
Figure 2.4.1 shows the risk of bias summary for studies investigating the effect of exercise on other behavioural outcomes in animals. The risk of bias assessment was performed using the SyRCLE RoB tool.
Figure 2.4.1
2.4.2 Reporting completeness
Figure 2.4.2 shows the reporting completeness summary for studies investigating the effect of exercise on other behavioural outcomes in animals. The reporting completeness assessment was performed using the ARRIVE guidelines.
Figure 2.4.2
2.4.3 Meta-analysis
The effect of exercise on other behavioural outcomes in animals using SMD as the effect size is shown in Figure 2.4.3. The pooled estimate for SMD across all individual comparisons is displayed as a diamond shape at the bottom of the plot. Dotted lines indicate the prediction interval of the pooled estimate.
Figure 2.4.3
Exercise interventions had a pooled effect on other behavioural outcomes induced by single prolonged stress of SMD = 1.287, (95% CI: 0.807 to 1.768; 95% PrI: -0.444 to 3.019).
74 experimental comparisons were reported in 19 experiments from 14 publications and involving 2 different animal strain(s).
| Level | Number of categories for that level included in this analysis | Attributable variance |
|---|---|---|
| Strain | 0 | NA |
| Study x Strain | 0 | NA |
| Study x Strain x Experiment | 0 | NA |
| Study | 14 | 0.238 |
| Study x Experiment | 19 | 0.355 |
2.4.4 Subgroup analyses and meta-regressions
For each outcome, the covariates of interest for subgroup analyses and meta-regressions were:
- Sex
- Voluntary or forced exercise
- Duration of exercise
- Intensity of exercise
We also conducted subgroup analyses using (1) SyRCLE Risk of Bias and (2) ARRIVE reporting completeness assessment scores as covariates to evaluate their influence on effect size estimates. These were not specified in the study protocol, but evaluation of risk of bias is required for the Summary of Evidence table, and no studies were considered entirely at low risk of bias or of high reporting completeness to allow such a sensitivity analysis.
The significance (p-value) reported is that for a test of whether the moderators are significantly different one from another, rather than whether the effect is significantly different from 0.
Sex
Figure 2.4.4.1 displays the estimates for the pooled SMD’s when comparisons are stratified by sex of the animal. Whiskers indicate the 95% confidence interval of each estimate. The overall pooled SMD, not stratified by sex, is displayed as a diamond shape at the bottom of the plot, with the 95% prediction interval shown as a red bar.
Figure 2.4.4.1 - Effect of exercise on other behavioural outcomes by Sex
The p-value for the association between the sex of animal groups used and outcome reported was 0.185.
| Level | Number of categories for that level included in this analysis | Attributable variance |
|---|---|---|
| Study | 14 | 0.492 |
| Study x Experiment | 19 | 0.191 |
Voluntary or forced exercise
Figure 2.4.4.2 displays the estimates for the pooled SMD when comparisons are stratified by whether the exercise was voluntary or forced (VoF). Whiskers indicate the 95% confidence interval of each estimate. The overall pooled SMD, not stratified by exercise type, is displayed as a diamond shape at the bottom of the plot, with the 95% prediction interval shown as a red bar.
Figure 2.4.4.2 - Effect of exercise on other behavioural outcomes by voluntary or forced exercise
The p-value for the association between the VoF of animal groups used and outcome reported was 0.18.
| Level | Number of categories for that level included in this analysis | Attributable variance |
|---|---|---|
| Study | 14 | 0.178 |
| Study x Experiment | 19 | 0.374 |
Duration of treatment
We provide a meta-regression of the number of weeks of treatment as a continuous variable.
Figure 2.4.4.3 - Effect of exercise on other behavioural outcomes by Duration of treatment
The p-value for the association between duration of treatment and outcome reported was 0.455.
| Level | Number of categories for that level included in this analysis | Attributable variance |
|---|---|---|
| Strain | 2 | 0 |
| Study x Strain | 14 | 0.25 |
| Study x Strain x Experiment | 19 | 0.383 |
Exercise Intensity
We provide a meta-regression where exercise intensity (expressed as m/min) is considered as a continuous variable.
Figure 2.4.4.4 - Effect of exercise on other behavioural outcomes by exercise intensity
The p-value for the association between exercise intensity and the outcome reported was 0.375.
| Level | Number of categories for that level included in this analysis | Attributable variance |
|---|---|---|
| Strain | 2 | 0.096 |
| Study x Strain | 14 | 0.254 |
| Study x Strain x Experiment | 19 | 0.369 |
Total exercise
We provide a meta-regression where the total exercise as the product of the number of sessions (n), session duration (min), and session intensity (km/min), with total exercise expressed in km (n * min * km/min), is considered as a continuous variable.
Figure 2.4.4.5
The p-value for the association between total exercise and the outcome reported was 0.003.
| Level | Number of categories for that level included in this analysis | Attributable variance |
|---|---|---|
| Strain | 2 | 0 |
| Study x Strain | 14 | 0.097 |
| Study x Strain x Experiment | 19 | 0.571 |
SyRCLE RoB assessment considered as a categorical variable
Figure 2.4.4.6 displays the estimates for the pooled SMD when comparisons are stratified by how many of the SyRCLE risk of bias assessment criteria (of which there are 10) the experiment met. Whiskers indicate the 95% confidence interval of each estimate. The overall pooled SMD, not stratified by SyRCLE Risk of Bias, is displayed as a diamond shape at the bottom of the plot, with the 95% prediction interval shown as a red bar.
Figure 2.4.4.6 - Effect of exercise on other behavioural outcomes by SyRCLE risk of bias criteria
The p-value for the association between SyRCLE Risks of Bias reporting and the outcome reported was 0.408.
| Level | Number of categories for that level included in this analysis | Attributable variance |
|---|---|---|
| Study | 14 | 0.284 |
| Study x Experiment | 19 | 0.338 |
SyRCLE RoB assessment considering those studies where any item is at low risk of bias
Figure 2.4.4.7 displays the estimates for the pooled SMD when comparisons are stratified by whether of not any of the SyRCLE Risk of bias domains were rated as low risk of bias. Whiskers indicate the 95% confidence interval of each estimate. The overall pooled SMD, not stratified by SyRCLE Risk of Bias, is displayed as a diamond shape at the bottom of the plot, with the 95% prediction interval shown as a red bar.
Figure 2.4.4.7 - Effect of exercise on other behavioural outcomes by low SyRCLE RoB
The p-value for the association between low SyRCLE Risks of Bias reporting and the outcome reported was 0.408.
| Level | Number of categories for that level included in this analysis | Attributable variance |
|---|---|---|
| Study | 14 | 0.284 |
| Study x Experiment | 19 | 0.338 |
ARRIVE reporting completeness guidelines
We provide a meta-regression where the number of ARRIVE items met is considered as a continuous variable.
Figure 2.4.4.8 - Effect of exercise on other behavioural outcomes by ARRIVE reporting completeness
The p-value for the association between ARRIVE reporting completeness and outcome reported was 0.848.
| Level | Number of categories for that level included in this analysis | Attributable variance |
|---|---|---|
| Strain | 2 | 0 |
| Study x Strain | 14 | 0.284 |
| Study x Strain x Experiment | 19 | 0.385 |
Heterogeneity explained by covariates (Effect of exercise on other behaviours)
The table below shows which of the covariates, if any, explain some of the heterogeneity observed in the effect of exercise on other behaviours. We present marginal R2, which measures the proportion of variance explained by including moderators in the model (the % change in the between-studies variance when the covariate is included in the model, in other words the % of the heterogeneity explained by the variable). The coefficients are derived form an RMA model fitted with an intercept (and so represent, for each category, the point estimate and 95% CIs of the effect in that category).
| Moderator | Category | \(\beta\) | 95% CI | Marginal R2 (%) |
|---|---|---|---|---|
| Overall effect | - | 1.287 | 0.807 to 1.768 | - |
| Sex | - | - | - | 8.2% |
| - | Female | 0.809 | -0.281 to 1.899 | - |
| - | Male | 1.365 | 0.339 to 2.39 | - |
| Voluntary or forced | - | - | - | 20.8% |
| - | Forced | 1.409 | 0.909 to 1.909 | - |
| - | Voluntary | 0.492 | -0.818 to 1.802 | - |
| Duration of treatment | - | - | - | 4.4% |
| - | per weeks of treatment increase | -0.143 | -0.545 to 0.26 | - |
| Exercise intensity | - | - | - | 0.9% |
| - | per unit (m/min) increase | -0.024 | -0.079 to 0.03 | - |
| Total exercise | - | - | - | 14.9% |
| - | per km increase | -0.06 | -0.098 to -0.022 | - |
| Risk of Bias | - | - | - | 5.9% |
| - | 0 criteria met | 1.211 | 0.682 to 1.741 | - |
| - | 1 criteria met | 1.813 | 0.376 to 3.25 | - |
| Reporting completeness | - | - | - | 0.3% |
| - | per unit increase | -0.024 | -0.285 to 0.238 | - |
2.4.5 Sensitivity Analyses
We examine the robustness of the findings for other behaviours by performing the following sensitivity analyses.
Imputed rho values of 0.2 and 0.8
In the previous analyses for the effect of exercise on other behavioural outcomes, we imputed a \(\rho\) value - the imputed within-study correlation between observed effect sizes - of 0.5. Here, we examine the effect of imputing \(\rho\) values of 0.2 and 0.8.
When the \(\rho\) value is assumed to be 0.2, exercise had a pooled effect on other behavioural outcomes of SMD = 1.54 (95% CI: 1.01 to 2.08; 95% PrI: -0.42 to 3.51).
When the \(\rho\) value is assumed to be 0.8, exercise had a pooled effect on other behavioural outcomes of SMD = 0.59 (95% CI: -0.08 to 1.25; 95% PrI: -1.98 to 3.15).
For reference the pooled effect size when rho is assumed to be 0.5 is 1.29 (95% CI: 0.81 to 1.77).
2.4.6 Reporting bias/small-study effects
Because of the relationship between SMD effect sizes and variance inherent in their calculation, where study size is small the standard approach to seeking evidence of small-study effects (regression based tests including Egger’s regression test for multilevel meta-analysis) can lead to over-estimation of small-study effect (see for instance 10.7554/eLife.24260). To address this we used Egger’s regression test for multilevel meta-analysis, with regression of SMD effect size against 1/√N, where N is the total number of animals involved in an experiment.
Egger regression based on 74 studies of modelling of depression where sucrose preference was measured showed a coefficient for a small-study effect of -1.48 (95% CI: -22.89 to 19.92; p = 0.882).
2.5 Outcome 5: BDNF
2.5.1 Risks of bias
Figure 2.5.1 shows the risk of bias summary for studies investigating the effect of exercise on BDNF in animals. The risk of bias assessment was performed using the SyRCLE RoB tool.
Figure 2.5.1
2.5.2 Reporting completeness
Figure 2.5.2 shows the reporting completeness summary for studies investigating the effect of exercise on BDNF in animals. The reporting completeness assessment was performed using the ARRIVE guidelines.
Figure 2.5.2
2.5.3 Meta-analysis
The effect of exercise on BDNF in animals using SMD as the effect size is shown in Figure 2.5.3. The pooled estimate for SMD across all individual comparisons is displayed as a diamond shape at the bottom of the plot. Dotted lines indicate the prediction interval of the pooled estimate.
Figure 2.5.3
Exercise had a pooled effect on BDNF of SMD = 1.79 , (95% CI: 0.56 to 3.01; 95% PrI: -1.63 to 5.21).
24 experimental comparisons were reported in 11 experiments from 8 publications and involving 2 different animal strain(s).
| Level | Number of categories for that level included in this analysis | Attributable variance |
|---|---|---|
| Study | 8 | 1.82 |
| Study x Experiment | 11 | 0 |
2.5.4 Subgroup analyses and meta-regressions
For each outcome, the covariates of interest for subgroup analyses and meta-regressions were:
- Sex
- Voluntary or forced exercise
- Duration of exercise
- Intensity of exercise
We also conducted subgroup analyses using (1) SyRCLE Risk of Bias and (2) ARRIVE reporting completeness assessment scores as covariates to evaluate their influence on effect size estimates. These were not specified in the study protocol, but evaluation of risk of bias is required for the Summary of Evidence table, and no studies were considered entirely at low risk of bias or of high reporting completeness to allow such a sensitivity analysis.
The significance (p-value) reported is that for a test of whether the moderators are significantly different one from another, rather than whether the effect is significantly different from 0.
Sex
Figure 2.5.4.1 displays the estimates for the pooled SMD when comparisons are stratified by sex of the animal. Whiskers indicate the 95% confidence interval of each estimate. The overall pooled SMD, not stratified by sex, is displayed as a diamond shape at the bottom of the plot, with the 95% prediction interval shown as a red bar.
Figure 2.5.4.1 - Effect of exercise on BDNF by Sex
The p-value for the association between the sex of animal groups used and outcome reported was 0.847.
| Level | Number of categories for that level included in this analysis | Attributable variance |
|---|---|---|
| Study | 8 | 1.892 |
| Study x Experiment | 11 | 0 |
Voluntary or forced exercise
Figure 2.5.4.2 displays the estimates for the pooled SMD when comparisons are stratified by whether the exercise was voluntary or forced (VoF). Whiskers indicate the 95% confidence interval of each estimate. The overall pooled SMD, not stratified by exercise type, is displayed as a diamond shape at the bottom of the plot, with the 95% prediction interval shown as a red bar.
Figure 2.5.4.2 - Effect of exercise on other BDNF by voluntary or forced exercise
The p-value for the association between the VoF of animal groups used and outcome reported was 0.868.
| Level | Number of categories for that level included in this analysis | Attributable variance |
|---|---|---|
| Study | 8 | 2.228 |
| Study x Experiment | 11 | 0 |
Duration of treatment
We provide a meta-regression of the number of weeks of treatment as a continuous variable.
Figure 2.5.4.3 - Effect of exercise on BDNF by duration of treatment
The p-value for the association between duration of treatment and outcome reported was 0.428.
| Level | Number of categories for that level included in this analysis | Attributable variance |
|---|---|---|
| Strain | 2 | 0 |
| Study x Strain | 8 | 1.992 |
| Study x Strain x Experiment | 11 | 0 |
Exercise Intensity
We provide a meta-regression where exercise intensity (expressed at m/min) is considered as a continuous variable.
Figure 2.5.4.4 - Effect of exercise on BDNF by exercise intensity
The p-value for the association between exercise intensity and outcome reported was 0.007.
| Level | Number of categories for that level included in this analysis | Attributable variance |
|---|---|---|
| Strain | 2 | 0 |
| Study x Strain | 8 | 2.811 |
| Study x Strain x Experiment | 11 | 0 |
Total exercise
We provide a meta-regression where the total exercise as the product of the number of sessions, session duration, and session intensity, expressed in km, is considered as a continuous variable.
Figure 2.5.4.5 - Effect of exercise on BDNF by total exercise
The p-value for the association between total exercise and outcome reported was 0.006.
| Level | Number of categories for that level included in this analysis | Attributable variance |
|---|---|---|
| Strain | 2 | 0 |
| Study x Strain | 8 | 2.711 |
| Study x Strain x Experiment | 11 | 0 |
SyRCLE RoB assessment considered as a categorical variable
No studies met any RoB criteria.
SyRCLE RoB assessment considering those studies where any item is at low risk of bias
No studies met any RoB criteria.
ARRIVE reporting completeness guidelines
We provide a meta-regression where the number of ARRIVE items met is considered as a continuous variable.
Figure 2.5.4.6 - Effect of exercise on BDNF by ARRIVE reporting completeness
The p-value for the association between ARRIVE reporting completeness and outcome reported was 0.804.
| Level | Number of categories for that level included in this analysis | Attributable variance |
|---|---|---|
| Strain | 2 | 0 |
| Study x Strain | 8 | 2.235 |
| Study x Strain x Experiment | 11 | 0 |
Heterogeneity explained by covariates (Effect of exercise on BDNF)
The table below shows which of the covariates, if any, explain some of the heterogeneity observed in the effect of exercise on BDNF. We present marginal R2, which measures the proportion of variance explained by including moderators in the model (the % change in the between-studies variance when the covariate is included in the model, in other words the % of the heterogeneity explained by the variable). The coefficients are derived form an RMA model fitted with an intercept (and so represent, for each category, the point estimate and 95% CIs of the effect in that category).
| Moderator | Category | \(\beta\) | 95% CI | Marginal R2 (%) |
|---|---|---|---|---|
| Overall effect | - | 1.789 | 0.564 to 3.013 | - |
| Sex | - | - | - | 0.1% |
| - | Female | 1.751 | 0.469 to 3.033 | - |
| - | Male | 1.832 | 0.559 to 3.106 | - |
| Voluntary or forced | - | - | - | 0.3% |
| - | Forced | 1.885 | -19.253 to 23.024 | - |
| - | Voluntary | 1.305 | -26.738 to 29.349 | - |
| Duration of treatment | - | - | - | 11.2% |
| - | per weeks of treatment increase | 0.369 | -0.695 to 1.433 | - |
| Exercise intensity | - | - | - | 8.7% |
| - | per unit (m/min) increase | -0.18 | -0.304 to -0.055 | - |
| Total exercise | - | - | - | 7.8% |
| - | per km increase | -0.077 | -0.129 to -0.024 | - |
| Reporting completeness | - | - | - | 0.9% |
| - | per unit increase | 0.076 | -0.642 to 0.794 | - |
2.5.5 Sensitivity Analyses
We examine the robustness of the findings for BDNF by performing the following sensitivity analyses.
Imputed rho values of 0.2 and 0.8
In the previous analyses for the effect of exercise on BDNF, we imputed a \(\rho\) value - the imputed within-study correlation between observed effect sizes - of 0.5. Here, we examine the effect of imputing \(\rho\) values of 0.2 and 0.8.
When the \(\rho\) value is assumed to be 0.2, exercise had a pooled effect on BDNF of SMD = 2.13 (95% CI: 0.64 to 3.62; 95% PrI: -2.15 to 6.4).
When the \(\rho\) value is assumed to be 0.8, exercise had a pooled effect on BDNF of SMD = 1.08 (95% CI: 0.04 to 2.12; 95% PrI: -1.81 to 3.97).
For reference the pooled effect size when rho is assumed to be 0.5 is 1.79 (95% CI: 0.56 to 3.01).
2.5.6 Reporting bias/small-study effects
Because of the relationship between SMD effect sizes and variance inherent in their calculation, where study size is small the standard approach to seeking evidence of small-study effects (regression based tests including Egger’s regression test for multilevel meta-analysis) can lead to over-estimation of small-study effect (see for instance 10.7554/eLife.24260). To address this we used Egger’s regression test for multilevel meta-analysis, with regression of SMD effect size against 1/√N, where N is the total number of animals involved in an experiment.
Egger regression based on 24 studies of modelling of depression where sucrose preference was measured showed a coefficient for a small-study effect of -5.94 (95% CI: -48.46 to 36.59; p = 0.744).
2.6 Outcome 6: Biological stress response
2.6.1 Risks of bias
Figure 2.6.1 shows the risk of bias summary for studies investigating the effect of exercise on biological stress response in animals. The risk of bias assessment was performed using the SyRCLE RoB tool.
Figure 2.6.1
2.6.2 Reporting completeness
Figure 2.6.2 shows the reporting completeness summary for studies investigating the effect of exercise on biological stress response in animals. The reporting completeness assessment was performed using the ARRIVE guidelines.
Figure 2.6.2
2.6.3 Meta-analysis
The effect of exercise on biological stress response in animals using SMD as the effect size is shown in Figure 2.6.3. The pooled estimate for SMD across all individual comparisons is displayed as a diamond shape at the bottom of the plot. Dotted lines indicate the prediction interval of the pooled estimate.
Figure 2.6.3
Exercise interventions had a pooled effect on biological stress response induced by single prolonged stress of SMD = 2.03, (95% CI: -1.78 to 5.84; 95% PrI: -7.88 to 11.93).
10 experimental comparisons were reported in 8 experiments from 6 publications and involving 1 different animal strain(s).
| Level | Number of categories for that level included in this analysis | Attributable variance |
|---|---|---|
| Study | 6 | 12.65 |
| Study x Experiment | 8 | 0 |
2.6.4 Subgroup analyses and meta-regressions
For each outcome, the covariates of interest for subgroup analyses and meta-regressions were:
- Sex
- Voluntary or forced exercise
- Duration of exercise
- Intensity of exercise
We also conducted subgroup analyses using (1) SyRCLE Risk of Bias and (2) ARRIVE reporting completeness assessment scores as covariates to evaluate their influence on effect size estimates. These were not specified in the study protocol, but evaluation of risk of bias is required for the Summary of Evidence table, and no studies were considered entirely at low risk of bias or of high reporting completeness to allow such a sensitivity analysis.
The significance (p-value) reported is that for a test of whether the moderators are significantly different one from another, rather than whether the effect is significantly different from 0.
Sex
Figure 2.6.4.1 displays the estimates for the pooled SMD when comparisons are stratified by sex of the animal. Whiskers indicate the 95% confidence interval of each estimate. The overall pooled SMD, not stratified by sex, is displayed as a diamond shape at the bottom of the plot, with the 95% prediction interval shown as a red bar.
Figure 2.6.4.1 - Effect of exercise on biological stress response by Sex
The p-value for the association between the sex of animal groups used and biological stress response was 0.631.
| Level | Number of categories for that level included in this analysis | Attributable variance |
|---|---|---|
| Study | 6 | 12.575 |
| Study x Experiment | 8 | 0 |
Voluntary or forced exercise
All studies used forced exercise.
Duration of treatment
We provide a meta-regression of the number of weeks of treatment as a continuous variable.
Figure 2.6.4.2 - Effect of exercise on biological stress response by duration of treatment
The p-value for the association between duration of treatment and biological stress response was 0.878.
| Level | Number of categories for that level included in this analysis | Attributable variance |
|---|---|---|
| Strain | 1 | 0 |
| Study x Strain | 6 | 16.169 |
| Study x Strain x Experiment | 8 | 0 |
Exercise Intensity
We provide a meta-regression where exercise intensity is considered as a continuous variable.
Figure 2.6.4.3 - Effect of exercise on biological stress response by exercise intensity
The p-value for the association between exercise intensity and biological stress response was 0.005.
| Level | Number of categories for that level included in this analysis | Attributable variance |
|---|---|---|
| Strain | 1 | 0 |
| Study x Strain | 6 | 12.248 |
| Study x Strain x Experiment | 8 | 0 |
Total exercise
We provide a meta-regression where the total exercise as the product of the number of sessions, session duration, and session intensity, expressed in km, is considered as a continuous variable.
Figure 2.6.4.4 - Effect of exercise on biological stress response by total exercise
The p-value for the association between total exercise and biological stress response was 0.006.
| Level | Number of categories for that level included in this analysis | Attributable variance |
|---|---|---|
| Strain | 1 | 0 |
| Study x Strain | 6 | 12.251 |
| Study x Strain x Experiment | 8 | 0 |
SyRCLE RoB assessment considered as a categorical variable
Figure 2.6.4.5 displays the estimates for the pooled SMD when comparisons are stratified by how many of the SyRCLE risk of bias assessment criteria (of which there are 10) that the experiment met. Whiskers indicate the 95% confidence interval of each estimate. The overall pooled SMD, not stratified by SyRCLE Risk of Bias, is displayed as a diamond shape at the bottom of the plot, with the 95% prediction interval shown as a red bar.
Figure 2.6.4.5 - Effect of exercise on biological stress response by SyRCLE RoB criteria
The p-value for the association between SyRCLE Risks of Bias reporting and biological stress response was 0.878.
| Level | Number of categories for that level included in this analysis | Attributable variance |
|---|---|---|
| Study | 6 | 16.169 |
| Study x Experiment | 8 | 0 |
SyRCLE RoB assessment considering those studies where any item is at low risk of bias
Figure 2.6.4.6 displays the estimates for the pooled SMD when comparisons are stratified by whether of not any of the SyRCLE Risk of bias domains were rated as low risk of bias. Whiskers indicate the 95% confidence interval of each estimate. The overall pooled SMD, not stratified by SyRCLE Risk of Bias, is displayed as a diamond shape at the bottom of the plot, with the 95% prediction interval shown as a red bar.
Figure 2.6.4.6 - Effect of exercise on biological stress response by low SyRCLE RoB
The p-value for the association between low SyRCLE Risks of Bias reporting and biological stress response was 0.878.
| Level | Number of categories for that level included in this analysis | Attributable variance |
|---|---|---|
| Study | 6 | 16.169 |
| Study x Experiment | 8 | 0 |
ARRIVE reporting completeness guidelines
We provide a meta-regression where the number of ARRIVE items met is considered as a continuous variable.
Figure 2.6.4.7 - Effect of exercise on biological stress response by ARRIVE reporting completeness
The p-value for the association between ARRIVE reporting completeness and biological stress response was 0.625.
| Level | Number of categories for that level included in this analysis | Attributable variance |
|---|---|---|
| Strain | 1 | 0 |
| Study x Strain | 6 | 15.235 |
| Study x Strain x Experiment | 8 | 0 |
Heterogeneity explained by covariates (Effect of exercise on biological stress response)
The table below shows which of the covariates, if any, explain some of the heterogeneity observed in the effect of exercise on biological stress response. We present marginal R2, which measures the proportion of variance explained by including moderators in the model (the % change in the between-studies variance when the covariate is included in the model, in other words the % of the heterogeneity explained by the variable). The coefficients are derived form an RMA model fitted with an intercept (and so represent, for each category, the point estimate and 95% CIs of the effect in that category).
| Moderator | Category | \(\beta\) | 95% CI | Marginal R2 (%) |
|---|---|---|---|---|
| Overall effect | - | 2.028 | -1.783 to 5.839 | - |
| Sex | - | - | - | 0.3% |
| - | Female | 2.23 | -1.518 to 5.978 | - |
| - | Male | 1.88 | -1.807 to 5.568 | - |
| Duration of treatment | - | - | - | 0.3% |
| - | per weeks of treatment increase | 0.364 | -5.81 to 6.537 | - |
| Exercise intensity | - | - | - | 4.4% |
| - | per unit (m/min) increase | -0.315 | -0.508 to -0.122 | - |
| Total exercise | - | - | - | 3% |
| - | per km increase | -0.132 | -0.212 to -0.051 | - |
| Risk of Bias | - | - | - | 0.3% |
| - | 0 criteria met | 2.166 | -2.92 to 7.252 | - |
| - | 1 criteria met | 1.439 | -9.812 to 12.69 | - |
| Reporting completeness | - | - | - | 4.9% |
| - | per unit increase | -0.408 | -2.551 to 1.736 | - |
2.6.5 Sensitivity Analyses
We examine the robustness of the findings for biological stress response by performing the following sensitivity analyses.
Imputed rho values of 0.2 and 0.8
In the previous analyses for the effect of exercise on biological stress response, we imputed a \(\rho\) value - the imputed within-study correlation between observed effect sizes - of 0.5. Here, we examine the effect of imputing \(\rho\) values of 0.2 and 0.8.
When the \(\rho\) value is assumed to be 0.2, exercise had a pooled effect on biological stress response of SMD = 2.04 (95% CI: -1.76 to 5.85; 95% PrI: -7.84 to 11.93).
When the \(\rho\) value is assumed to be 0.8, exercise had a pooled effect on biological stress response of SMD = 1.96 (95% CI: -1.88 to 5.81; 95% PrI: -8.04 to 11.96).
For reference the pooled effect size when rho is assumed to be 0.5 is 2.03 (95% CI: -1.78 to 5.84).
2.6.6 Reporting bias/small-study effects
Because of the relationship between SMD effect sizes and variance inherent in their calculation, where study size is small the standard approach to seeking evidence of small-study effects (regression based tests including Egger’s regression test for multilevel meta-analysis) can lead to over-estimation of small-study effect (see for instance 10.7554/eLife.24260). To address this we used Egger’s regression test for multilevel meta-analysis, with regression of SMD effect size against 1/√N, where N is the total number of animals involved in an experiment.
Egger regression based on 10 studies of modelling of depression where other neurobiological outcomes was measured showed a coefficient for a small-study effect of -111.34 (95% CI: -271.68 to 48.99; p = 0.126).
2.7 Outcome 7: Neurotransmitter levels
Multilevel analysis is only performed if there are 5 levels or more for at least one of Strain, Study and Experiment, and that is not the case here. 10 experimental comparisons were reported in 1 experiments from 1 publications and involving 1 different animal strain(s). We provide a conventional random effects model to illustrate the data. No subgroup analysis is performed. We also now include these Freezing outcomes under the ‘other neurobiological’ heading below.
2.8 Outcome 8: Other neurobiological outcomes
2.8.1 Risks of bias
Figure 2.8.1 shows the risk of bias summary for studies investigating the effect of exercise on other neurobiological outcomes in animals. The risk of bias assessment was performed using the SyRCLE RoB tool.
Figure 2.8.1
2.8.2 Reporting completeness
Figure 2.8.2 shows the reporting completeness summary for studies investigating the effect of exercise on other neurobiological outcomes in animals. The reporting completeness assessment was performed using the ARRIVE guidelines.
Figure 2.8.2
2.8.3 Meta-analysis
The effect of exercise on other neurobiological outcomes in animals using SMD as the effect size is shown in Figure 2.8.3. The pooled estimate for SMD across all individual comparisons is displayed as a diamond shape at the bottom of the plot. Dotted lines indicate the prediction interval of the pooled estimate.
Figure 2.8.3
Exercise interventions had a pooled effect on other neurobiological outcomes induced by single prolonged stress of SMD = 1.1, (95% CI: -0.18 to 2.38; 95% PrI: -2.79 to 4.99).
62 experimental comparisons were reported in 11 experiments from 9 publications and involving 2 different animal strain(s).
| Level | Number of categories for that level included in this analysis | Attributable variance |
|---|---|---|
| Study | 9 | 2.53 |
| Study x Experiment | 11 | 0 |
2.8.4 Subgroup analyses and meta-regressions
For each outcome, the covariates of interest for subgroup analyses and meta-regressions were:
- Sex
- Voluntary or forced exercise
- Duration of exercise
- Intensity of exercise
- Control control conditions
We also conducted subgroup analyses using (1) SyRCLE Risk of Bias and (2) ARRIVE reporting completeness assessment scores as covariates to evaluate their influence on effect size estimates. These were not specified in the study protocol, but evaluation of risk of bias is required for the Summary of Evidence table, and no studies were considered entirely at low risk of bias or of high reporting completeness to allow such a sensitivity analysis.
The significance (p-value) reported is that for a test of whether the moderators are significantly different one from another, rather than whether the effect is significantly different from 0.
Sex
Figure 2.8.4.1 displays the estimates for the pooled SMD when comparisons are stratified by sex of the animal. Whiskers indicate the 95% confidence interval of each estimate. The overall pooled SMD, not stratified by sex, is displayed as a diamond shape at the bottom of the plot, with the 95% prediction interval shown as a red bar.
Figure 2.8.4.1 - Effect of exercise on other neurobiological outcomes by Sex
The p-value for the association between the sex of animal groups used and other neurobiological outcomes was 0.935.
| Level | Number of categories for that level included in this analysis | Attributable variance |
|---|---|---|
| Study | 9 | 2.677 |
| Study x Experiment | 11 | 0 |
Voluntary or forced exercise
Figure 2.8.4.2 displays the estimates for the pooled SMD when comparisons are stratified by whether the exercise was voluntary or forced. Whiskers indicate the 95% confidence interval of each estimate. The overall pooled SMD, not stratified by sex, is displayed as a diamond shape at the bottom of the plot, with the 95% prediction interval shown as a red bar.
Figure 2.8.4.2 - Effect of exercise on other neurobiological outcomes by voluntary or forced exercise
The p-value for the association between the VoF of animal groups used and other neurobiological outcomes was 0.181.
| Level | Number of categories for that level included in this analysis | Attributable variance |
|---|---|---|
| Study | 9 | 2.146 |
| Study x Experiment | 11 | 0 |
Duration of treatment
We provide a meta-regression of the number of weeks of treatment as a continuous variable.
Figure 2.8.4.3 - Effect of exercise on other neurobiological outcomes by duration of treatment
The p-value for the association between duration of treatment and other neurobiological outcomes was 0.291.
| Level | Number of categories for that level included in this analysis | Attributable variance |
|---|---|---|
| Strain | 2 | 0 |
| Study x Strain | 9 | 2.444 |
| Study x Strain x Experiment | 11 | 0 |
Exercise Intensity
We provide a meta-regression where exercise intensity is considered as a continuous variable.
Figure 2.8.4.4 - Effect of exercise on other neurobiological outcomes by exercise intensity
The p-value for the association between exercise intensity and outcome reported was 0.04.
| Level | Number of categories for that level included in this analysis | Attributable variance |
|---|---|---|
| Strain | 2 | 0 |
| Study x Strain | 9 | 2.56 |
| Study x Strain x Experiment | 11 | 0 |
Total exercise
We provide a meta-regression where the total exercise as the product of the number of sessions, session duration, and session intensity, expressed in km, is considered as a continuous variable.
Figure 2.8.4.5 - Effect of exercise on other neurobiological outcomes by total exercise
The p-value for the association between total exercise and other neurobiological outcomes was 0.063.
| Level | Number of categories for that level included in this analysis | Attributable variance |
|---|---|---|
| Strain | 2 | 0 |
| Study x Strain | 9 | 2.906 |
| Study x Strain x Experiment | 11 | 0 |
SyRCLE RoB assessment considered as a categorical variable
No studies met any RoB criteria.
SyRCLE RoB assessment considering those studies where any item is at low risk of bias
No studies met any RoB criteria.
ARRIVE reporting completeness guidelines
We provide a meta-regression where the number of ARRIVE items met is considered as a continuous variable.
Figure 2.8.4.6 - Effect of exercise on other neurobiological outcomes by ARRIVE reporting completeness
The p-value for the association between ARRIVE reporting completeness and other neurobiological outcomes was 0.223.
| Level | Number of categories for that level included in this analysis | Attributable variance |
|---|---|---|
| Strain | 2 | 0 |
| Study x Strain | 9 | 2.398 |
| Study x Strain x Experiment | 11 | 0 |
Heterogeneity explained by covariates (Effect of exercise on other neurobiological outcomes)
The table below shows which of the covariates, if any, explain some of the heterogeneity observed in the effect of exercise on other neurobiological outcomes. We present marginal R2, which measures the proportion of variance explained by including moderators in the model (the % change in the between-studies variance when the covariate is included in the model, in other words the % of the heterogeneity explained by the variable). The coefficients are derived form an RMA model fitted with an intercept (and so represent, for each category, the point estimate and 95% CIs of the effect in that category).
| Moderator | Category | \(\beta\) | 95% CI | Marginal R2 (%) |
|---|---|---|---|---|
| Overall effect | - | 1.101 | -0.181 to 2.384 | - |
| Sex | - | - | - | 0% |
| - | Female | 1.133 | -0.375 to 2.642 | - |
| - | Male | 1.074 | -0.451 to 2.599 | - |
| Voluntary or forced | - | - | - | 28% |
| - | Forced | 1.513 | 0.12 to 2.906 | - |
| - | Voluntary | -0.305 | -2.839 to 2.228 | - |
| Duration of treatment | - | - | - | 26% |
| - | per weeks of treatment increase | -0.983 | -3.021 to 1.054 | - |
| Exercise intensity | - | - | - | 2.2% |
| - | per unit (m/min) increase | 0.055 | 0.002 to 0.108 | - |
| Total exercise | - | - | - | 2.7% |
| - | per km increase | 0.054 | -0.003 to 0.112 | - |
| Reporting completeness | - | - | - | 13.8% |
| - | per unit increase | -0.34 | -0.94 to 0.261 | - |
2.8.5 Sensitivity Analyses
We examine the robustness of the findings for the primary outcome by performing the following sensitivity analyses.
Imputed rho values of 0.2 and 0.8
In the previous analyses for the effect of exercise on other neurobiological outcomes, we imputed a \(\rho\) value - the imputed within-study correlation between observed effect sizes - of 0.5. Here, we examine the effect of imputing \(\rho\) values of 0.2 and 0.8.
When the \(\rho\) value is assumed to be 0.2, exercise had a pooled effect on other neurobiological outcomes of SMD = 1.69 (95% CI: 0.62 to 2.76; 95% PrI: -1.53 to 4.91).
When the \(\rho\) value is assumed to be 0.8, exercise had a pooled effect on other neurobiological outcomes of SMD = 0.09 (95% CI: -1.73 to 1.91; 95% PrI: -5.54 to 5.72).
For reference the pooled effect size when rho is assumed to be 0.5 is 1.1 (95% CI: -0.18 to 2.38.
2.8.6 Reporting bias/small-study effects
Because of the relationship between SMD effect sizes and variance inherent in their calculation, where study size is small the standard approach to seeking evidence of small-study effects (regression based tests including Egger’s regression test for multilevel meta-analysis) can lead to over-estimation of small-study effect (see for instance 10.7554/eLife.24260). To address this we used Egger’s regression test for multilevel meta-analysis, with regression of SMD effect size against 1/√N, where N is the total number of animals involved in an experiment.
Egger regression based on 62 studies of modelling of depression where other neurobiological outcomes was measured showed a coefficient for a small-study effect of -27.2 (95% CI: -62.64 to 8.24; p = 0.112).
3 Effects of model induction (SPS)
14 studies (143 comparisons) investigated the effects of model induction. The number of studies and individual effect sizes for each outcome were:
Locomotor outcomes: 4 studies and 13 comparisons in 2 strains
Fear Memory 4 studies and 7 comparisons in 1 strain
Freezing: 3 studies and 3 comparisons in 2 strains
Other behaviours: 11 studies and 41 comparisons in 2 strains
BDNF: 8 studies and 20 comparisons in 2 strains
Neurotransmitter levels: 1 studies and 10 comparisons in 1 strain
Biochemical stress responses: 6 studies and 9 comparisons in 1 strain
Other neurobiological outcomes: 8 studies and 40 comparisons in 2 strains
3.1 Outcome 1: Locomotor activity
Multilevel analysis is only performed if there are 5 levels or more for at least one of Strain, Study and Experiment, and that is not the case here. 13 experimental comparisons were reported in 4 experiments from 4 publications and involving 2 different animal strain(s). We provide a conventional random effects model to illustrate the data. No subgroup analysis is performed. We also now include these locomotor outcomes under the ‘other behavioural’ heading below.
3.2 Outcome 2: Fear memory
Multilevel analysis is only performed if there are 5 levels or more for at least one of Strain, Study and Experiment, and that is not the case here. 7 experimental comparisons were reported in 4 experiments from 4 publications and involving 1 different animal strain(s). We provide a conventional random effects model to illustrate the data. No subgroup analysis is performed. We also now include these fear memory outcomes under the ‘other behavioural’ heading below.
3.3 Outcome 3: Freezing
Multilevel analysis is only performed if there are 5 levels or more for at least one of Strain, Study and Experiment, and that is not the case here. 3 experimental comparisons were reported in 3 experiments from 3 publications and involving 2 different animal strain(s). We provide a conventional random effects model to illustrate the data. No subgroup analysis is performed. We also now include these freezing outcomes under the ‘other behavioural’ heading below.
3.4 Outcome 4: Other behavioural outcomes
3.4.1 Risks of bias
Figure 3.4.1 shows the risk of bias summary for studies investigating the effect of single prolonged stress on other behavioural outcomes in animals. The risk of bias assessment was performed using the SyRCLE RoB tool.
Figure 3.4.1
3.4.2 Reporting completeness
Figure 3.4.2 shows the reporting completeness summary for studies investigating the effect of single prolonged stress on sucrose preference in animals. The reporting completeness assessment was performed using the ARRIVE guidelines.
Figure 3.4.2
3.4.3 Meta-analysis
The effect of single prolonged stress on other behaviours in animals using SMD as the effect size is shown in Figure 3.4.3. The pooled estimate for SMD across all individual comparisons is displayed as a diamond shape at the bottom of the plot. Dotted lines indicate the prediction interval of the pooled estimate.
Figure 3.4.3
Single prolonged stress had a pooled effect on other behaviours of SMD = -1.64, (95% CI: -2.5 to -0.78; 95% PrI: -4.79 to 1.5).
64 experimental comparisons were reported in 19 experiments from 14 publications and involving 2 different animal strain(s).
The following table structure is used throughout this report and is used to show the different levels contributing to that analysis, the number of unique categories in those levels, and the variance contributed by that level of analysis. Because levels are only included in the analysis where there are five or more unique categories, for some analyses the number of categories is 0, and the variance attributed to those levels in not applicable. Because the model is hierarchical, where for instance there are Studies which include different Strains, the number of categories for Study x Strain will exceed the number of Studies (by which we mean unique publications) referred to in the text.
| Level | Number of categories for that level included in this analysis | Attributable variance |
|---|---|---|
| Study | 14 | 1.88 |
| Study x Experiment | 19 | 0.08 |
3.4.4 Subgroup analyses and meta-regressions
For each outcome, the covariates of interest for subgroup analyses and meta-regressions were:
- Sex
We also conducted subgroup analyses using (1) SyRCLE Risk of Bias and (2) ARRIVE reporting completeness assessment scores as covariates to evaluate their influence on effect size estimates. These were not specified in the study protocol, but evaluation of risk of bias is required for the Summary of Evidence table, and no studies were considered entirely at low risk of bias or of high reporting completeness to allow such a sensitivity analysis
The significance (p-value) reported is that for a test of whether the moderators are significantly different one from another, rather than whether the effect is significantly different from 0.
Sex
Figure 3.4.4.1 displays the estimates for the pooled SMD’s when comparisons are stratified by sex of the animal. Whiskers indicate the 95% confidence interval of each estimate. The overall pooled SMD, not stratified by sex, is displayed as a diamond shape at the bottom of the plot, with the 95% prediction interval shown as a red bar.
Figure 3.4.4.1 - Effect of single prolonged stress on other behavioural outcomes by Sex
The p-value for the association between the sex of animal groups used and outcome reported was 0.225.
| Level | Number of categories for that level included in this analysis | Attributable variance |
|---|---|---|
| Study | 14 | 2.218 |
| Study x Experiment | 19 | 0.001 |
SyRCLE RoB assessment considered as a categorical variable
Figure 3.4.4.3 displays the estimates for the pooled SMD’s when comparisons are stratified by how many of the SyRCLE risk of bias assessment criteria (of which there are 10) that the experiment met. Whiskers indicate the 95% confidence interval of each estimate. The overall pooled SMD, not stratified by SyRCLE Risk of Bias, is displayed as a diamond shape at the bottom of the plot, with the 95% prediction interval shown as a red bar.
Figure 3.4.4.3 - Effect of single prolonged stress on other behavioural outcomes by SyRCLE RoB criteria
The p-value for the association between SyRCLE Risks of Bias reporting and outcome reported was 0.134.
| Level | Number of categories for that level included in this analysis | Attributable variance |
|---|---|---|
| Study | 14 | 2.018 |
| Study x Experiment | 19 | 0.083 |
SyRCLE RoB assessment considering those studies where any item is at low risk of bias
Figure 3.4.4.4 displays the estimates for the pooled SMD’s when comparisons are stratified by whether of not any of the SyRCLE Risk of bias domains were rated as low risk of bias. Whiskers indicate the 95% confidence interval of each estimate. The overall pooled SMD, not stratified by SyRCLE Risk of Bias, is displayed as a diamond shape at the bottom of the plot.
Figure 3.4.4.4 - Effect of single prolonged stress on other behavioural outcomes by low SyRCLE RoB
The p-value for the association between low SyRCLE Risks of Bias reporting and outcome reported was 0.134.
| Level | Number of categories for that level included in this analysis | Attributable variance |
|---|---|---|
| Study | 14 | 2.018 |
| Study x Experiment | 19 | 0.083 |
ARRIVE reporting completeness guidelines
We provide a meta-regression where the number of ARRIVE items met is considered as a continuous variable.
Figure 3.4.4.5 - Effect of single prolonged stress on other behavioural outcomes by ARRIVE reporting completeness
The p-value for the association between ARRIVE reporting completeness and outcome reported was 0.423.
| Level | Number of categories for that level included in this analysis | Attributable variance |
|---|---|---|
| Strain | 2 | 0 |
| Study x Strain | 14 | 2.194 |
| Study x Strain x Experiment | 19 | 0.082 |
Heterogeneity explained by covariates (Effect of modelling on Behaviour not otherwise specified)
The table below shows which of the covariates, if any, explain some of the heterogeneity observed in the effect of model induction on other behavioural outcomes. We present marginal R2, which measures the proportion of variance explained by including moderators in the model (the % change in the between-studies variance when the covariate is included in the model, in other words the % of the heterogeneity explained by the variable). The coefficients are derived form an RMA model fitted with an intercept (and so represent, for each category, the point estimate and 95% CIs of the effect in that category).
| Moderator | Category | \(\beta\) | 95% CI | Marginal R2 (%) |
|---|---|---|---|---|
| Overall effect | - | -1.643 | -2.502 to -0.784 | - |
| Sex | - | - | - | 1.9% |
| - | Female | -1.366 | -2.496 to -0.235 | - |
| - | Male | -1.81 | -2.902 to -0.719 | - |
| Risk of Bias | - | - | - | 19.6% |
| - | 0 criteria met | -1.412 | -2.363 to -0.46 | - |
| - | 1 criteria met | -3.456 | -6.059 to -0.853 | - |
| Reporting completeness | - | - | - | 4.5% |
| - | per unit increase | 0.214 | -0.348 to 0.777 | - |
3.4.5 Sensitivity Analyses
We examine the robustness of the findings for other behaviours by performing the following sensitivity analyses.
Imputed rho values of 0.2 and 0.8
In the previous analyses for the effect of single prolonged stress on other behaviours, we imputed a \(\rho\) value - the imputed within-study correlation between observed effect sizes - of 0.5. Here, we examine the effect of imputing \(\rho\) values of 0.2 and 0.8.
When the \(\rho\) value is assumed to be 0.2, single prolonged stress had a pooled effect on other behaviours of SMD = -1.89 (95% CI: -2.65 to -1.13; 95% PrI: -4.68 to 0.9).
When the \(\rho\) value is assumed to be 0.8, single prolonged stress had a pooled effect on other behaviours of SMD = -1.05 (95% CI: -2.23 to 0.12; 95% PrI: -5.5 to 3.39).
For reference the pooled effect size when rho is assumed to be 0.5 is -1.64 (95% CI: -2.5 to -0.78).
3.4.6 Reporting bias/small-study effects
Because of the relationship between SMD effect sizes and variance inherent in their calculation, where study size is small the standard approach to seeking evidence of small-study effects (regression based tests including Egger’s regression test for multilevel meta-analysis) can lead to over-estimation of small-study effect (see for instance 10.7554/eLife.24260). To address this we used Egger’s regression test for multilevel meta-analysis, with regression of SMD effect size against 1/√N, where N is the total number of animals involved in an experiment.
Egger regression based on 64 studies of modelling of depression where sucrose preference was measured showed a coefficient for a small-study effect of -6.44 (95% CI: -44.57 to 31.7; p = 0.719).
3.5 Outcome 5: BDNF
3.5.1 Risks of bias
Figure 3.5.1 shows the risk of bias summary for studies investigating the effect of single prolonged stress on sucrose preference in animals. The risk of bias assessment was performed using the SyRCLE RoB tool.
Figure 3.5.1
3.5.2 Reporting completeness
Figure 3.5.2 shows the reporting completeness summary for studies investigating the effect of single prolonged stress on sucrose preference in animals. The reporting completeness assessment was performed using the ARRIVE guidelines.
Figure 3.5.2
3.5.3 Meta-analysis
The effect of single prolonged stress on other behaviours in animals using SMD as the effect size is shown in Figure 3.5.3. The pooled estimate for SMD across all individual comparisons is displayed as a diamond shape at the bottom of the plot. Dotted lines indicate the prediction interval of the pooled estimate.
Figure 3.5.3
Single prolonged stress had a pooled effect on other behaviours of SMD = -2.11, (95% CI: -4.56 to 0.34; 95% PrI: -9.3 to 5.08).
20 experimental comparisons were reported in 11 experiments from 8 publications and involving 2 different animal strain(s).
| Level | Number of categories for that level included in this analysis | Attributable variance |
|---|---|---|
| Study | 8 | 8.16 |
| Study x Experiment | 11 | 0 |
3.5.4 Subgroup analyses and meta-regressions
For each outcome, the covariates of interest for subgroup analyses and meta-regressions were:
- Sex
We also conducted subgroup analyses using (1) SyRCLE Risk of Bias and (2) ARRIVE reporting completeness assessment scores as covariates to evaluate their influence on effect size estimates. These were not specified in the study protocol, but evaluation of risk of bias is required for the Summary of Evidence table, and no studies were considered entirely at low risk of bias or of high reporting completeness to allow such a sensitivity analysis
The significance (p-value) reported is that for a test of whether the moderators are significantly different one from another, rather than whether the effect is significantly different from 0.
Sex
Figure 3.5.4.1 displays the estimates for the pooled SMD’s when comparisons are stratified by sex of the animal. Whiskers indicate the 95% confidence interval of each estimate. The overall pooled SMD, not stratified by sex, is displayed as a diamond shape at the bottom of the plot, with the 95% prediction interval shown as a red bar.
Figure 3.5.4.1 - Effect of single prolonged stress on BDNF by Sex
The p-value for the association between the sex of animal groups used and outcome reported was 0.335.
| Level | Number of categories for that level included in this analysis | Attributable variance |
|---|---|---|
| Study | 8 | 7.935 |
| Study x Experiment | 11 | 0 |
SyRCLE RoB assessment considered as a categorical variable
No studies met any RoB criteria
SyRCLE RoB assessment considering those studies where any item is at low risk of bias
No studies met any RoB criteria
ARRIVE reporting completeness guidelines
We provide a meta-regression where the number of ARRIVE items met is considered as a continuous variable.
Figure 3.5.4.2 - Effect of single prolonged stress on BDNF by ARRIVE reporting completeness
The p-value for the association between ARRIVE reporting completeness and outcome reported was 0.61.
| Level | Number of categories for that level included in this analysis | Attributable variance |
|---|---|---|
| Strain | 2 | 0 |
| Study x Strain | 8 | 9.352 |
| Study x Strain x Experiment | 11 | 0 |
Heterogeneity explained by covariates (Effect of modelling on BDNF)
The table below shows which of the covariates, if any, explain some of the heterogeneity observed in the effect of modelling on BDNF. We present marginal R2, which measures the proportion of variance explained by including moderators in the model (the % change in the between-studies variance when the covariate is included in the model, in other words the % of the heterogeneity explained by the variable). The coefficients are derived form an RMA model fitted with an intercept (and so represent, for each category, the point estimate and 95% CIs of the effect in that category).
| Moderator | Category | \(\beta\) | 95% CI | Marginal R2 (%) |
|---|---|---|---|---|
| Overall effect | - | -2.111 | -4.561 to 0.339 | - |
| Sex | - | - | - | 0.6% |
| - | Female | -1.869 | -4.241 to 0.503 | - |
| - | Male | -2.301 | -4.652 to 0.051 | - |
| Reporting completeness | - | - | - | 3.4% |
| - | per unit increase | 0.382 | -1.357 to 2.121 | - |
3.5.5 Sensitivity Analyses
We examine the robustness of the findings for BDNF by performing the following sensitivity analyses.
Imputed rho values of 0.2 and 0.8
In the previous analyses for the effect of single prolonged stress on BDNF, we imputed a \(\rho\) value - the imputed within-study correlation between observed effect sizes - of 0.5. Here, we examine the effect of imputing \(\rho\) values of 0.2 and 0.8.
When the \(\rho\) value is assumed to be 0.2, single prolonged stress had a pooled effect on BDNF of SMD = -2.48 (95% CI: -5.13 to 0.17; 95% PrI: -10.3 to 5.33).
When the \(\rho\) value is assumed to be 0.8, single prolonged stress had a pooled effect on BDNF of SMD = -1.39 (95% CI: -3.57 to 0.79; 95% PrI: -7.83 to 5.04).
For reference the pooled effect size when rho is assumed to be 0.5 is -2.11 (95% CI: -4.56 to 0.34).
3.5.6 Reporting bias/small-study effects
Because of the relationship between SMD effect sizes and variance inherent in their calculation, where study size is small the standard approach to seeking evidence of small-study effects (regression based tests including Egger’s regression test for multilevel meta-analysis) can lead to over-estimation of small-study effect (see for instance 10.7554/eLife.24260). To address this we used Egger’s regression test for multilevel meta-analysis, with regression of SMD effect size against 1/√N, where N is the total number of animals involved in an experiment.
Egger regression based on 20 studies of modelling of depression where sucrose preference was measured showed a coefficient for a small-study effect of 36.06 (95% CI: -41.43 to 113.55; p = 0.298).
3.6 Outcome 6: Biological stress response
3.6.1 Risks of bias
Figure 3.6.1 shows the risk of bias summary for studies investigating the effect of model induction on biological stress response in animals. The risk of bias assessment was performed using the SyRCLE RoB tool.
Figure 3.6.1
3.6.2 Reporting completeness
Figure 3.6.2 shows the reporting completeness summary for studies investigating the effect of model induction on biological stress response in animals. The reporting completeness assessment was performed using the ARRIVE guidelines.
Figure 3.6.2
3.6.3 Meta-analysis
The effect of single prolonged stress on the biological stress response in animals using SMD as the effect size is shown in Figure 3.6.3. The pooled estimate for SMD across all individual comparisons is displayed as a diamond shape at the bottom of the plot. Dotted lines indicate the prediction interval of the pooled estimate.
Figure 3.6.3
Single prolonged stress had a pooled effect on biological stress response of SMD = -3.41, (95% CI: -5.32 to -1.51; 95% PrI: -8.13 to 1.3).
9 experimental comparisons were reported in 8 experiments from 6 publications and involving 1 different animal strain(s).
| Level | Number of categories for that level included in this analysis | Attributable variance |
|---|---|---|
| Study | 6 | 2.82 |
| Study x Experiment | 8 | 0 |
3.6.4 Subgroup analyses and meta-regressions
For each outcome, the covariates of interest for subgroup analyses and meta-regressions were:
- Sex
We also conducted subgroup analyses using (1) SyRCLE Risk of Bias and (2) ARRIVE reporting completeness assessment scores as covariates to evaluate their influence on effect size estimates. These were not specified in the study protocol, but evaluation of risk of bias is required for the Summary of Evidence table, and no studies were considered entirely at low risk of bias or of high reporting completeness to allow such a sensitivity analysis.
The significance (p-value) reported is that for a test of whether the moderators are significantly different one from another, rather than whether the effect is significantly different from 0.
Sex
Figure 3.6.4.1 displays the estimates for the pooled SMD’s when comparisons are stratified by sex of the animal. Whiskers indicate the 95% confidence interval of each estimate. The overall pooled SMD, not stratified by sex, is displayed as a diamond shape at the bottom of the plot, with the 95% prediction interval shown as a red bar.
Figure 3.6.4.1 - Effect of single prolonged stress on biological stress response by Sex
The p-value for the association between the sex of animal groups used and outcome reported was 0.932.
| Level | Number of categories for that level included in this analysis | Attributable variance |
|---|---|---|
| Study | 6 | 2.557 |
| Study x Experiment | 8 | 0.501 |
SyRCLE RoB assessment considered as a categorical variable
Figure 3.6.4.2 displays the estimates for the pooled SMD when comparisons are stratified by how many of the SyRCLE risk of bias assessment criteria (of which there are 10) that the experiment met. Whiskers indicate the 95% confidence interval of each estimate. The overall pooled SMD, not stratified by SyRCLE Risk of Bias, is displayed as a diamond shape at the bottom of the plot, with the 95% prediction interval shown as a red bar.
Figure 3.6.4.2 - Effect of single prolonged stress on biological stress response by SyRCLE RoB
The p-value for the association between SyRCLE Risks of Bias reporting and outcome reported was 0.414.
| Level | Number of categories for that level included in this analysis | Attributable variance |
|---|---|---|
| Study | 6 | 2.893 |
| Study x Experiment | 8 | 0 |
SyRCLE RoB assessment considering those studies where any item is at low risk of bias
Figure 3.6.4.3 displays the estimates for the pooled SMD when comparisons are stratified by whether of not any of the SyRCLE Risk of bias domains were rated as low risk of bias. Whiskers indicate the 95% confidence interval of each estimate. The overall pooled SMD, not stratified by SyRCLE Risk of Bias, is displayed as a diamond shape at the bottom of the plot, with the 95% prediction interval shown as a red bar.
Figure 3.6.4.3 - Effect of single prolonged stress on biological stress response by low SyRCLE RoB
The p-value for the association between low SyRCLE Risks of Bias reporting and outcome reported was 0.414.
| Level | Number of categories for that level included in this analysis | Attributable variance |
|---|---|---|
| Study | 6 | 2.893 |
| Study x Experiment | 8 | 0 |
ARRIVE reporting completeness guidelines
We provide a meta-regression where the number of ARRIVE items met is considered as a continuous variable.
Figure 3.6.4.4 - Effect of single prolonged stress on biological stress response by ARRIVE reporting completeness
The p-value for the association between ARRIVE reporting completeness and outcome reported was 0.688.
| Level | Number of categories for that level included in this analysis | Attributable variance |
|---|---|---|
| Strain | 1 | 0 |
| Study x Strain | 6 | 3.419 |
| Study x Strain x Experiment | 8 | 0 |
Heterogeneity explained by covariates (Effect of modelling on Biological stress response)
The table below shows which of the covariates, if any, explain some of the heterogeneity observed in the effect of model induction on biological stress response. We present marginal R2, which measures the proportion of variance explained by including moderators in the model (the % change in the between-studies variance when the covariate is included in the model, in other words the % of the heterogeneity explained by the variable). The coefficients are derived form an RMA model fitted with an intercept (and so represent, for each category, the point estimate and 95% CIs of the effect in that category).
| Moderator | Category | \(\beta\) | 95% CI | Marginal R2 (%) |
|---|---|---|---|---|
| Overall effect | - | -3.414 | -5.319 to -1.509 | - |
| Sex | - | - | - | 0.1% |
| - | Female | -3.407 | -5.907 to -0.908 | - |
| - | Male | -3.5 | -5.551 to -1.449 | - |
| Risk of Bias | - | - | - | 12.3% |
| - | 0 criteria met | -3.128 | -5.386 to -0.869 | - |
| - | 1 criteria met | -5.038 | -10.398 to 0.323 | - |
| Reporting completeness | - | - | - | 3.2% |
| - | per unit increase | 0.203 | -1.098 to 1.503 | - |
3.6.5 Sensitivity Analyses
We examine the robustness of the findings for biological stress response by performing the following sensitivity analyses.
Imputed rho values of 0.2 and 0.8
In the previous analyses for the effect of single prolonged stress on biological stress response, we imputed a \(\rho\) value - the imputed within-study correlation between observed effect sizes - of 0.5. Here, we examine the effect of imputing \(\rho\) values of 0.2 and 0.8.
When the \(\rho\) value is assumed to be 0.2, single prolonged stress had a pooled effect on biological stress response of SMD = -3.27 (95% CI: -5.44 to -1.1; 95% PrI: -8.73 to 2.19).
When the \(\rho\) value is assumed to be 0.8, single prolonged stress had a pooled effect on biological stress response of SMD = -3.78 (95% CI: -5.31 to -2.26; 95% PrI: -7.43 to -0.14).
For reference the pooled effect size when rho is assumed to be 0.5 is -3.41 (95% CI: -5.32 to -1.51).
3.6.6 Reporting bias/small-study effects
Because of the relationship between SMD effect sizes and variance inherent in their calculation, where study size is small the standard approach to seeking evidence of small-study effects (regression based tests including Egger’s regression test for multilevel meta-analysis) can lead to over-estimation of small-study effect (see for instance 10.7554/eLife.24260). To address this we used Egger’s regression test for multilevel meta-analysis, with regression of SMD effect size against 1/√N, where N is the total number of animals involved in an experiment.
Egger regression based on 9 studies of modelling of depression where sucrose preference was measured showed a coefficient for a small-study effect of 22.53 (95% CI: -84.41 to 129.47; p = 0.590).
3.7 Outcome 7: Neurotransmitter levels
Multilevel analysis is only performed if there are 5 levels or more for at least one of Strain, Study and Experiment, and that is not the case here. 10 experimental comparisons were reported in 1 experiments from 1 publications and involving 1 different animal strain(s). We provide a conventional random effects model to illustrate the data. No subgroup analysis is performed. We also now include these freezing outcomes under the ‘other neurobiological’ heading below.
3.8 Outcome 8: Other neurobiological outcomes
3.8.1 Risks of bias
Figure 3.8.1 shows the risk of bias summary for studies investigating the effect of single prolonged stress on other neurobiological outcomes in animals. The risk of bias assessment was performed using the SyRCLE RoB tool.
Figure 3.8.1
3.8.2 Reporting completeness
Figure 3.8.2 shows the reporting completeness summary for studies investigating the effect of single prolonged stress on other neurobiological outcomes in animals. The reporting completeness assessment was performed using the ARRIVE guidelines.
Figure 3.8.2
3.8.3 Meta-analysis
The effect of single prolonged stress on other neurobiological outcomes in animals using SMD as the effect size is shown in Figure 3.1.3. The pooled estimate for SMD across all individual comparisons is displayed as a diamond shape at the bottom of the plot. Dotted lines indicate the prediction interval of the pooled estimate.
Figure 3.8.3
Single prolonged stress had a pooled effect on other neurobiological outcomes of SMD = -1.68, (95% CI: -3.82 to 0.46; 95% PrI: -7.92 to 4.56).
50 experimental comparisons were reported in 9 experiments from 8 publications and involving 2 different animal strain(s).
| Level | Number of categories for that level included in this analysis | Attributable variance |
|---|---|---|
| Study | 8 | 6.14 |
| Study x Experiment | 9 | 0 |
3.8.4 Subgroup analyses and meta-regressions
For each outcome, the covariates of interest for subgroup analyses and meta-regressions were:
- Sex
We also conducted subgroup analyses using (1) SyRCLE Risk of Bias and (2) ARRIVE reporting completeness assessment scores as covariates to evaluate their influence on effect size estimates. These were not specified in the study protocol, but evaluation of risk of bias is required for the Summary of Evidence table, and no studies were considered entirely at low risk of bias or of high reporting completeness to allow such a sensitivity analysis.
The significance (p-value) reported is that for a test of whether the moderators are significantly different one from another, rather than whether the effect is significantly different from 0.
Sex
Figure 3.8.4.1 displays the estimates for the pooled SMD when comparisons are stratified by sex of the animal. Whiskers indicate the 95% confidence interval of each estimate. The overall pooled SMD, not stratified by sex, is displayed as a diamond shape at the bottom of the plot, with the 95% prediction interval shown as a red bar.
Figure 3.8.4.1 - Effect of single prolonged stress on other neurobiological outcomes by Sex
The p-value for the association between the sex of animal groups used and outcome reported was 0.335.
| Level | Number of categories for that level included in this analysis | Attributable variance |
|---|---|---|
| Study | 8 | 6.539 |
| Study x Experiment | 9 | 0 |
SyRCLE RoB assessment considered as a categorical variable
No studies met any RoB criteria.
SyRCLE RoB assessment considering those studies where any item is at low risk of bias
No studies met any RoB criteria.
ARRIVE reporting completeness guidelines
We provide a meta-regression where the number of ARRIVE items met is considered as a continuous variable.
Figure 3.8.4.2 - Effect of single prolonged stress on other neurobiological outcomes by ARRIVE reporting completeness
The p-value for the association between ARRIVE reporting completeness and outcome reported was 0.233.
| Level | Number of categories for that level included in this analysis | Attributable variance |
|---|---|---|
| Strain | 2 | 0 |
| Study x Strain | 8 | 6.001 |
| Study x Strain x Experiment | 9 | 0 |
Heterogeneity explained by covariates (Effect of modelling on other neurobiological outcomes)
The table below shows which of the covariates, if any, explain some of the heterogeneity observed in the effect of modelling on other neurobiological outcomes. We present marginal R2, which measures the proportion of variance explained by including moderators in the model (the % change in the between-studies variance when the covariate is included in the model, in other words the % of the heterogeneity explained by the variable). The coefficients are derived form an RMA model fitted with an intercept (and so represent, for each category, the point estimate and 95% CIs of the effect in that category).
| Moderator | Category | \(\beta\) | 95% CI | Marginal R2 (%) |
|---|---|---|---|---|
| Overall effect | - | -1.681 | -3.821 to 0.46 | - |
| Sex | - | - | - | 0.4% |
| - | Female | -1.864 | -4.256 to 0.527 | - |
| - | Male | -1.554 | -3.878 to 0.771 | - |
| Reporting completeness | - | - | - | 14.6% |
| - | per unit increase | 0.652 | -0.551 to 1.856 | - |
3.8.5 Sensitivity Analyses
We examine the robustness of the findings for other neurobiological outcomes by performing the following sensitivity analyses
Imputed rho values of 0.2 and 0.8
In the previous analyses for the effect of single prolonged stress on other neurobiological outcomes, we imputed a \(\rho\) value - the imputed within-study correlation between observed effect sizes - of 0.5. Here, we examine the effect of imputing \(\rho\) values of 0.2 and 0.8.
When the \(\rho\) value is assumed to be 0.2, single prolonged stress had a pooled effect on other neurobiological outcomes of SMD = -2.38 (95% CI: -4.29 to -0.47; 95% PrI: -7.94 to 3.18).
When the \(\rho\) value is assumed to be 0.8, single prolonged stress had a pooled effect on other neurobiological outomes of SMD = -0.53 (95% CI: -3.2 to 2.13; 95% PrI: -8.38 to 7.32).
For reference the pooled effect size when rho is assumed to be 0.5 is -1.68 (95% CI: -3.82 to 0.46).
3.8.6 Reporting bias/small-study effects
Because of the relationship between SMD effect sizes and variance inherent in their calculation, where study size is small the standard approach to seeking evidence of small-study effects (regression based tests including Egger’s regression test for multilevel meta-analysis) can lead to over-estimation of small-study effect (see for instance 10.7554/eLife.24260). To address this we used Egger’s regression test for multilevel meta-analysis, with regression of SMD effect size against 1/√N, where N is the total number of animals involved in an experiment.
Egger regression based on 50 studies of modelling of depression where other neurobiological outcomes was measured showed a coefficient for a small-study effect of 26.02 (95% CI: -39.67 to 91.72; p = 0.370).
4. Observed relationships between different outcomes measures in the same cohorts of animals
We selected cohorts where at least one outcome was presented for at least two outcome types. Where there were two or more of the same outcome type within a cohort, we calculated a standardised mean difference effect size for that outcome in that cohort, along with its standard error. Where there was a single effect size within a cohort, we took the standard error of that effect size.
Then, for each pair of outcome measures we plotted the effect sizes for each cohort, and fitted a regression line weighted on the standard error in the outcome measure represented on the x-axis. Outcome measure pairs are coded according to whether they come from model induction studies (red, expectation of worsening anhedonia) or from intervention studies (green, expectation of improvement in anhedonia). The number of experimental comparisons observed from each cohort is reflected in the size of the symbol, and shown in the figure legend.
We show pairwise relationships where there were at least 3 experimental cohorts for either modelling experiments, intervention experiments, or both, and provide the relevant regression coefficient(s). We also provide an overall representation of all paired neurobiological effect sizes with the corresponding behavioural outcome in that cohort.
4.1 Relationship between change in locomotor activity and change in ‘other’ neurobiological outcomes
4.2 Relationship between change in freezing behaviour and change in ‘other’ behavioural outcomes
4.3 Relationship between change in BDNF and change in ‘other’ behavioural outcomes
4.4 Relationship between change in stress response and change in ‘other’ behavioural outcomes
4.5 Relationship between change in ‘other’ neurobiological outcomes and change in ‘other’ behavioural outcomes
4.6 Relationship between change in stress response and change in BDNF
4.7 Relationship between change in ‘other’ neurobiological outcomes and change in BDNF
4.8 Relationship between change in ‘other’ neurobiological outcomes and change in stress response
4.9 Relationship between change in all neurobiological outcomes and all behavioural outcomes
5. Attrition bias and adverse effects of treatment
We know of 1823 animals in Control cohorts and 1823 animals in Intervention cohorts, of which 0% and 0% were reported to have ‘dropped out’ between allocation to group and outcome measurement. This analysis assumes full reporting of animals excluded from analyses, and it may be that group sizes were specified ‘after the event’, or that there was unreported replacement of animals excluded during the experiment, so these data should be interpreted with caution.
6. Summary of the evidence
7. Software used
We used R version 4.3.1 (R Core Team 2023) and the following R packages: devtools v. 2.4.5 (Wickham et al. 2022), dosresmeta v. 2.0.1 (Crippa and Orsini 2016), ggpubr v. 0.6.0 (Kassambara 2023), gtools v. 3.9.5 (Warnes et al. 2023), Hmisc v. 5.1.1 (Harrell Jr 2023a), kableExtra v. 1.4.0.3 (Zhu 2024), knitr v. 1.45 (Xie 2014, 2015, 2023), Matrix v. 1.6.5 (Bates, Maechler, and Jagan 2024), meta v. 7.0.0 (Balduzzi, Rücker, and Schwarzer 2019), metadat v. 1.2.0 (White et al. 2022), metafor v. 4.4.0 (Viechtbauer 2010), mvmeta v. 1.0.3 (Gasparrini, Armstrong, and Kenward 2012), numDeriv v. 2016.8.1.1 (Gilbert and Varadhan 2019), orchaRd v. 2.0 (Nakagawa et al. 2023), patchwork v. 1.2.0 (Pedersen 2024), PRISMA2020 v. 1.1.1 (Haddaway et al. 2022), rje v. 1.12.1 (Evans 2022), rms v. 6.7.1 (Harrell Jr 2023b), robvis v. 0.3.0.900 (McGuinness and Higgins 2020), scales v. 1.3.0 (Wickham, Pedersen, and Seidel 2023), tidyverse v. 2.0.0 (Wickham et al. 2019), usethis v. 2.2.3 (Wickham et al. 2024), xtable v. 1.8.4 (Dahl et al. 2019).